LaRocca/qwen2.5-3b-medieval-npc-gguf

Qwen2.5-3B Medieval NPC — GGUF (Q4KM)

A medieval genre-pack fine-tune of Qwen2.5-3B-Instruct for game NPC dialogue, packaged for on-device, offline inference. It keeps a small model in medieval register, deflects out-of-era topics in character, and answers genuine pleas with in-character help — with zero per-conversation cost.

Built for the [local-npc](https://github.com/) runtime (a self-contained llama.cpp-based C core with Unity/Unreal bindings), but it is a standard GGUF and loads in any llama.cpp-compatible loader.

What it does

• —In-register everywhere — folksy medieval speech ("ye/aye/naught"), even with a bare persona card. The behaviour lives in the weights, not the prompt.
• —Deflects modern topics in character ("I know naught of such things…") instead of leaking assistant behaviour or discussing anachronisms.
• —Genuine pleas get in-character help — "my child has a fever" → concern + a period remedy; "someone stole my goat, my children are hungry" → practical, in-world aid — not an anti-assistant rebuff.

It is deliberately constrained: it is not a general assistant and will not reliably answer modern/factual questions — that is the point. Pair it with a persona card (character) and a deterministic term-guard (hard bans) in the local-npc runtime.

Model details

Usage (llama.cpp)

llama-cli -m qwen2.5-3b-medieval-q4_k_m.gguf -p "You are the elder of a small village." -cnv

In local-npc, point a persona card at this model — the pack supplies the register, the card supplies the character, and the runtime's term-guard is the hard, deterministic ban on named modern entities.

License & attribution

Licensed under Apache-2.0, inherited from the base model Qwen2.5-3B-Instruct (© Alibaba Cloud / Qwen team, Apache-2.0). This is a derivative work; you may use and redistribute it commercially under Apache-2.0, including inside shipped games. Retain this attribution.

Inference engine (not included here): llama.cpp — MIT.

Limitations

• —One adversarial phrasing ("can I charge my phone") can still fumble the wording; pair with the runtime term-guard for a hard guarantee.
• —English only; folksy Middle-English-flavoured register, not strict period philology.
• —Not a factual or general-purpose assistant by design.